High-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture

Abstract

The present invention discloses a high-computing-power SAR real-time imaging and target recognition system based on an FPGA+GPU architecture, which relates to the technical field of SAR imaging and target recognition. The system comprises: a SAR echo data acquisition module that collects multi-band data through a 16-bit AD / DA daughter card and an FMC interface; a preprocessing module that performs noise reduction, filtering, and multi-dimensional feature fusion based on an FPGA; a high-speed interaction module that uses a GTH and ETH interface to implement FPGA-GPU data transmission; a knowledge graph module that dynamically updates features and relationships; an imaging matching module that relies on a GPU to generate high-resolution images and output candidate categories; an inference optimization module that outputs accurate results and feeds back to the knowledge graph; the present invention implements preprocessing and feature extraction of SAR echo data through an FPGA chip, and utilizes its parallel processing capability to perform noise reduction, filtering, and range compression on the raw data, thereby improving data processing efficiency.

Description

Technical Field

[0001] The present invention relates to the field of SAR imaging and target recognition technology, and specifically to a high-computing-power SAR real-time imaging and target recognition system based on an FPGA+GPU architecture. Background Art

[0002] With the widespread application of synthetic aperture radar (SAR) technology in military reconnaissance, disaster monitoring, and resource exploration, higher requirements are placed on the real-time imaging and target recognition capabilities of SAR systems. Traditional SAR systems are limited by data processing speed and algorithm complexity, making it difficult to complete large-scale echo data processing and high-resolution image generation in a short period of time. Especially in dynamic scenarios, lack of timeliness has become a key bottleneck restricting its application. In recent years, the collaborative architecture of FPGA and GPU has provided new ideas for solving this problem. FPGA, with its parallel processing capability and low latency characteristics, is suitable for high-speed data preprocessing and feature extraction; GPU, with its powerful parallel computing capability, supports the real-time operation of complex imaging algorithms and target recognition models. The combination of the two provides a technical foundation for building high-computing-power, low-latency SAR real-time imaging systems, and promotes the development of SAR technology towards high efficiency and intelligence.

[0003] Traditional SAR systems have significant defects in data processing and target recognition. First, systems that rely solely on CPU processing are limited by serial computing modes and cannot cope with the high-concurrency processing requirements of large-scale SAR echo data, resulting in long imaging cycles and poor real-time performance. Second, traditional feature extraction methods mostly rely on single-dimensional features and ignore the fusion of multi-dimensional features, resulting in low target recognition accuracy, especially in complex scenarios. Misjudgment is prone to occur. In addition, traditional knowledge bases are mostly statically constructed and lack a dynamic update mechanism. They cannot promptly absorb the features and relationship information in new recognition results, resulting in insufficient timeliness of the knowledge base and difficulty in adapting to environmental changes. Finally, traditional systems lack efficient transmission protocols for data interaction. Data transmission between FPGA and GPU can easily become a performance bottleneck, further limiting the overall efficiency of the system. These problems jointly restrict the application potential of traditional SAR systems in real-time monitoring and dynamic decision-making. Summary of the Invention

[0004] The purpose of this invention is to make up for the shortcomings of the existing technology. It uses FPGA to realize high-speed data preprocessing and multi-dimensional feature fusion extraction, and uses GPU to accelerate high-resolution imaging algorithms and knowledge graph probabilistic reasoning, thereby improving data processing efficiency and target recognition accuracy. The system supports dynamic polarization mode switching and dynamic updating of knowledge graphs, adapts to complex scene changes, and solves the problems of poor real-time performance and low recognition accuracy of traditional SAR systems.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: a high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture, the system comprising:

[0006] SAR echo data acquisition module: This module uses a 16-bit AD / DA daughter card and FMC interface to collect multi-band SAR echo data, and transmits it to the data preprocessing and feature extraction module via the PCIe bus.

[0007] Data preprocessing and feature extraction module: Based on the FPGA chip, it performs noise reduction, filtering, and distance compression preprocessing on the original data, and simultaneously extracts and fuses the features of texture, contour, and scattering through a multi-dimensional feature fusion extraction algorithm;

[0008] High-speed data interaction module: connects the data transmission channel between FPGA and GPU, adopts the combination of GTH high-speed interface and ETH Ethernet interface, and combines data buffer queue, priority scheduling, check retransmission and flow control mechanism to realize the bidirectional transmission of feature data, control instructions and status information;

[0009] Knowledge graph storage and update module: Builds a knowledge graph library based on SAR target feature attributes, category labels, and inter-target relationship data. It also supports dynamic addition of features and relationships based on new recognition results, and regularly cleans outdated information.

[0010] Real-time imaging and feature matching module: Relying on GPU chips and parallel computing frameworks, it generates high-resolution SAR images using real-time imaging algorithms for received pre-processed data. It also synchronously calls the knowledge graph feature library, performs multi-dimensional feature matching, and uses a multimodal matching algorithm to compare the underlying features with the template to output candidate target categories.

[0011] Association reasoning and result optimization module: Relying on GPU computing power and integrating knowledge graph relationship rules, the initial matching results are deeply processed by the knowledge graph probabilistic reasoning model to output accurate recognition results containing category, location, and confidence, and at the same time, high-confidence results are fed back to the knowledge graph.

[0012] Furthermore, in the SAR echo data acquisition module, the AD / DA subcard supports dynamic switching of polarization modes, and realizes adaptive acquisition of HH, VV, HV, and VH polarization modes through real-time adjustment of the FPGA configuration register, and the multi-band switching response time is ≤100μs. When collecting data, parity check and CRC check are performed synchronously, and a unique check code is generated in combination with the timestamp.

[0013] Furthermore, in the data preprocessing and feature extraction module, the features of texture, contour, and scattering are extracted and fused by a multi-dimensional feature fusion extraction algorithm. The calculation formula of the multi-dimensional feature fusion extraction algorithm is: ,in is the polarization characteristic operator, is the scattering characteristic operator, is the texture feature operator, is a dynamic weight, which is adaptively controlled by the scene complexity calculated in real time by FPGA, and , For the moment The multi-dimensional comprehensive features after extraction and fusion, Indicates time.

[0014] Furthermore, in the knowledge graph storage and update module, the specific steps of constructing a knowledge graph based on SAR target feature attributes, category labels, and inter-target relationship data are as follows:

[0015] (1) Data normalization: receiving SAR target characteristics, categories, and relationship data and standardizing the format;

[0016] (2) Entity point construction: Based on the category label, the target is abstracted into an entity node and associated with feature attributes;

[0017] (3) Relationship Edges: Analyze the relationship between targets, build directed edges between entity nodes, and mark the relationship type;

[0018] (4) Knowledge fusion: merge duplicate entities and resolve relationship conflicts;

[0019] (5) Rule deduction: Inferring implicit relationships based on existing knowledge and extracting high-frequency pattern generation rules;

[0020] (6) Storage index: Store entities, relationships, and rules in a graph database format, establish classification indexes, and precipitate metadata.

[0021] Furthermore, in the knowledge graph storage and update module, features and relationships are dynamically supplemented. When the recognition result confidence level is ≥90%, the feature supplement process is triggered: the multi-dimensional feature vector of the newly recognized target is extracted. , Represent the characteristic components of different dimensions of the target, and calculate The similarity with each template in the existing feature library is as follows: , represent and The quantitative results of the similarity calculated between is the similarity calculation function, It represents the newly collected and newly input feature data used to participate in similarity calculation. It is the pre-stored feature data used to compare with the new feature. , then Add it to the library as a new feature template; when the recognition result reveals a new relationship, trigger the relationship supplement process: calculate the new relationship The frequency of occurrence of is: ,when hour, For this relationship The number of occurrences, Target The total number of related relationships will be Added to the knowledge base as a new relationship rule, Is a new relationship, Rel is a target object The related relationship function, is the target object in the SAR image.

[0022] Furthermore, in the knowledge graph storage and update module, the confidence calculation formula is: ,in is the target The recognition confidence of is the posterior probability of knowledge graph association reasoning, is the similarity score between the target feature and the knowledge graph template, with a value range of , is the scene consistency score, is the dynamic weight coefficient, and , determined by the complexity of the current scene, Represents the SAR image to be solved.

[0023] Furthermore, in the real-time imaging and feature matching module, a real-time imaging algorithm is used to generate a high-resolution SAR image. The calculation formula of the real-time imaging algorithm is: ,in yes High-resolution SAR images generated at all times, is a random sampling matrix, yes The square of the norm, It is the echo data after FPGA preprocessing. is the total variation regularization term, is the regularization coefficient.

[0024] Furthermore, in the real-time imaging and feature matching module, a multimodal matching algorithm is used to compare the underlying features with the template and output the target candidate category. The calculation formula of the multimodal matching algorithm is: , is the matching score between the target feature and the template, is the underlying feature vector to be matched, is the target template feature in the knowledge graph, is a multimodal feature matching operator, is the association weight between the feature and the knowledge graph, represents the knowledge graph, is the distance between the feature and the template, It is a distance-based attenuation factor calculated for all target templates in the knowledge graph. , filter the top scorers The category of the item is used as a candidate category .

[0025] Furthermore, in the association reasoning and result optimization module, the construction of the knowledge graph probability reasoning model is calculated as follows: ,in: is the optimal target category for final reasoning, is a candidate category, It is the initial matching probability based on the underlying features, It is a set of relationship rules related to the target in the knowledge graph. is the rule weight, which is dynamically adjusted by the rule confidence and scene complexity. is the target and rules degree of correlation.

[0026] Compared with existing technologies, this high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture has the following beneficial effects:

[0027] 1. This invention uses an FPGA chip to implement preprocessing and feature extraction of SAR echo data, leveraging its parallel processing capabilities to perform noise reduction, filtering, and range compression on the raw data, thereby improving data processing efficiency. Furthermore, the FPGA supports a multi-dimensional feature fusion algorithm with dynamic weight control, adaptively adjusting feature extraction strategies based on scene complexity to ensure the comprehensiveness and accuracy of feature information. The introduction of a GPU enables real-time generation of high-resolution SAR images. Combined with a parallel computing framework to optimize imaging algorithms, the system can complete large-scale data processing and imaging tasks in a short period of time.

[0028] 2. The present invention realizes efficient screening and accurate identification of target candidate categories through GPU-accelerated multimodal matching algorithm and knowledge graph probabilistic reasoning model. The multimodal matching algorithm integrates multi-dimensional features of texture, contour, and scattering, and combines the target template and relationship rules in the knowledge graph to output candidate categories with higher confidence; and the knowledge graph probabilistic reasoning model further utilizes relationship rules and scene context to deeply optimize the preliminary results and output accurate results including category, location and confidence. In addition, the system supports dynamic updating of knowledge graphs. When the confidence of the recognition result exceeds the threshold, new features and relationship rules are automatically added, and outdated information is regularly cleaned to ensure the timeliness and accuracy of the knowledge base.

[0029] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0031] Figure 1 This is a flowchart of a high-computing SAR real-time imaging and target recognition system based on FPGA+GPU architecture;

[0032] Figure 2 This is the module architecture diagram of the SAR real-time imaging and target recognition system based on FPGA+GPU architecture. DETAILED DESCRIPTION

[0033] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0034] Example 1: Application of a high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture in military target monitoring in urban areas.

[0035] SAR echo data acquisition module:

[0036] Through the 16-bit AD / DA daughter card and FMC interface, multi-band SAR echo data is collected for urban military target monitoring scenarios. Due to the large differences in the electromagnetic scattering characteristics of buildings, vegetation, and military equipment targets in cities, the AD / DA daughter card dynamically switches the polarization mode through the FPGA configuration register and adaptively collects echo data in four polarization modes: HH, VV, HV, and VH, to more comprehensively capture the scattering characteristics of different targets. Parity check and CRC check are performed synchronously during the collection process, and a unique check code is generated in combination with the timestamp to ensure the accuracy of data transmission. The data is then transmitted to the data preprocessing and feature extraction module via the PCIe bus. Figure 1 shown.

[0037] Data preprocessing and feature extraction module:

[0038] Based on the FPGA chip, the received raw echo data is pre-processed by noise reduction, filtering, and range compression to remove interference signals (such as building reflection clutter and electromagnetic noise) in the complex electromagnetic environment of the city, thereby improving data quality. At the same time, through the multi-dimensional feature fusion extraction algorithm, the texture features (such as equipment surface texture), contour features (such as equipment outline), and scattering features (such as the target's reflection intensity distribution of electromagnetic waves) of military targets (such as armored vehicles and radar stations) are extracted from the pre-processed data, and the three features are fused into a comprehensive feature. The calculation formula of the multi-dimensional feature fusion extraction algorithm is as follows: ,in is the polarization characteristic operator, is the scattering characteristic operator, is the texture feature operator, For the moment The multi-dimensional comprehensive features after extraction and fusion, Indicates time, is a dynamic weight, which is adaptively controlled by the scene complexity calculated in real time by FPGA, and Provide a basis for subsequent identification.

[0039] High-speed data interaction module:

[0040] By combining the GTH high-speed interface with the ETH Ethernet interface, a high-speed data transmission channel is established between the FPGA and GPU. To ensure efficient data and instruction transmission, the module uses a data buffer queue, priority scheduling, verification retransmission, and flow control mechanisms to transmit the fused feature data, control instructions (such as imaging parameter adjustment instructions), and status information (such as data processing progress) output by the FPGA to the GPU. At the same time, it receives control signals fed back by the GPU to achieve two-way real-time interaction.

[0041] Knowledge graph storage and update module:

[0042] A knowledge graph is constructed based on existing urban military target feature attributes (such as the typical texture and outline of armored vehicles), category labels (such as "armored vehicle" and "radar station"), and target relationship data (such as "radar stations are often adjacent to command vehicles"). During the monitoring process, this module continuously receives new recognition results. If new military target features or relationships between targets are found, they will be dynamically added to the knowledge graph. When the confidence level of the recognition result is ≥90%, the feature supplement process is triggered: the multi-dimensional feature vector of the newly recognized target is extracted. , Represent the characteristic components of different dimensions of the target, and calculate The similarity with each template in the existing feature library is as follows: , represent and The quantitative results of the similarity calculated between is the similarity calculation function, It represents the newly collected and newly input feature data used to participate in similarity calculation. It is the pre-stored feature data used to compare with the new feature. , then It is added to the library as a new feature template, and the confidence calculation formula is: ,in is the target The recognition confidence of is the posterior probability of knowledge graph association reasoning, is the similarity score between the target feature and the knowledge graph template, with a value range of , is the scene consistency score, is the dynamic weight coefficient, and , which is determined by the complexity of the current scene; when the recognition results reveal new relationships, the relationship supplement process is triggered: calculate the new relationship The frequency of occurrence of is: ,when hour, For this relationship The number of occurrences, Target The total number of related relationships will be Added to the knowledge base as a new relationship rule, Is a new relationship, Rel is a target object The related relationship function, It is the target object in the SAR image; at the same time, outdated information (such as features of obsolete equipment) is regularly cleaned to ensure the timeliness and accuracy of the knowledge graph.

[0043] Real-time imaging and feature matching module:

[0044] Relying on GPU chips and a parallel computing framework, a real-time imaging algorithm is applied to the received pre-processed data to generate high-resolution SAR images of urban areas, clearly showing the location and shape of buildings, roads, and military targets. The calculation formula of the real-time imaging algorithm is: ,in yes High-resolution SAR images generated at all times, is a random sampling matrix, yes The square of the norm, It is the echo data after FPGA preprocessing. is the total variation regularization term, represents the SAR image to be solved, is the regularization coefficient, and the knowledge graph feature library is called synchronously. The extracted comprehensive features of military targets are compared with the templates in the knowledge graph (such as the feature templates of known armored vehicles and radar stations) through the multimodal matching algorithm, and the target candidate categories (such as "armored vehicles", "tanks", and "civilian vehicles") are output. The calculation formula of the multimodal matching algorithm is: , is the matching score between the target feature and the template, is the underlying feature vector to be matched, is the target template feature in the knowledge graph, is a multimodal feature matching operator, is the association weight between the feature and the knowledge graph, represents the knowledge graph, is the distance between the feature and the template, It is a distance-based attenuation factor calculated for all target templates in the knowledge graph. , filter the top scorers The category of the item is used as a candidate category .

[0045] Association reasoning and result optimization module:

[0046] Relying on GPU computing power, the system integrates relational rules in the knowledge graph (such as "armored vehicles are often found around military bases") and deeply processes the candidate categories that are initially matched using the knowledge graph probabilistic reasoning model. For example, if a target is initially matched as an "armored vehicle" and the knowledge graph shows that there are "military base"-related targets near this location, the reasoning model will use this relationship to increase the confidence level of the "armored vehicle" category. The final output is an accurate recognition result that includes the target category (such as "armored vehicle"), location (such as a street coordinate), and confidence level. The calculation formula of the knowledge graph probabilistic reasoning model is: ,in: is the optimal target category for final reasoning, is a candidate category, It is the initial matching probability based on the underlying features, It is a set of relationship rules related to the target in the knowledge graph. is the rule weight, which is dynamically adjusted by the rule confidence and scene complexity. is the target and rules and feeds back high-confidence results (such as the “radar station” recognition result with a confidence level ≥ 90%) to the knowledge graph for updating the feature library and relationship rules.

[0047] In summary, in the monitoring of military targets in urban areas, the system obtains high-quality raw data through the SAR echo data acquisition module, obtains effective features through preprocessing and feature extraction, the high-speed data interaction module ensures data transmission, the knowledge graph provides knowledge support, real-time imaging and feature matching narrow the recognition range, and associative reasoning optimizes the results and feeds back to the knowledge graph. The entire process is efficient and coordinated, and with the help of the functions and algorithms of each module, accurate and real-time recognition of military targets is achieved.

[0048] Example 2: Application of a high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture in marine ship monitoring.

[0049] SAR echo data acquisition module:

[0050] Through the 16-bit AD / DA daughter card and FMC interface, multi-band SAR echo data is collected for marine ship monitoring scenarios. Due to the different polarization characteristics of waves, islands, and ships in the marine environment (such as the difference in reflection polarization between metal ships and seawater), the AD / DA daughter card adaptively switches the HH, VV, HV, and VH polarization modes through the FPGA configuration register to quickly respond to the monitoring needs of different sea areas. Parity check and CRC check are performed synchronously during collection, and a unique check code is generated in combination with the timestamp to ensure that the echo data (such as the electromagnetic wave signal reflected by the ship) is accurate. It is then transmitted to the data preprocessing and feature extraction module through the PCIe bus, such as Figure 2 shown.

[0051] Data preprocessing and feature extraction module:

[0052] Based on the FPGA chip, the original echo data is pre-processed by noise reduction, filtering, and range compression to remove wave clutter and atmospheric interference noise, highlighting the signal of the ship target. At the same time, the multi-dimensional feature fusion extraction algorithm is used to extract the ship's texture features (such as the surface structure texture of the hull), contour features (such as the length and width profile of the hull), and scattering features (such as the reflection intensity distribution of the hull to electromagnetic waves) and fuse them into comprehensive features. The calculation formula of the multi-dimensional feature fusion extraction algorithm is: , distinguish between ships, islands and floating objects.

[0053] High-speed data interaction module:

[0054] By combining the GTH high-speed interface with the ETH Ethernet interface, a high-speed data transmission channel between the FPGA and the GPU is constructed. The module is equipped with a data buffer queue, priority scheduling, verification retransmission, and flow control mechanism to transmit the ship fusion feature data, control instructions (such as imaging range adjustment instructions), and status information (such as data transmission status) output by the FPGA to the GPU. At the same time, it receives instructions fed back by the GPU to ensure efficient two-way transmission of data and instructions.

[0055] Knowledge graph storage and update module:

[0056] A knowledge graph is constructed based on existing marine vessel attribute features (such as the outline of a freighter, the scattering characteristics of an oil tanker), category labels (such as "freighter," "oil tanker," and "fishing boat"), and inter-target relationship data (such as "oil tankers are often adjacent to ports"). During the monitoring process, if a new ship type (such as a new type of container ship) or a new relationship (such as "container ships are often related to freight terminals") is identified, the module will add the new features and relationships to the knowledge graph. When the confidence level of the recognition result is ≥90%, the feature supplementation process is triggered: the multi-dimensional feature vector of the newly identified target is extracted. ,calculate The similarity with each template in the existing feature library is as follows: , when all , then Add it to the library as a new feature template; when the recognition result reveals a new relationship, trigger the relationship supplement process: calculate the new relationship The frequency of occurrence of is: ,when hour, For this relationship The number of occurrences, Target The total number of related relationships will be Add new relationship rules to the knowledge base; at the same time, regularly clean outdated information (such as the characteristics of old ships) to maintain the validity of the knowledge graph.

[0057] Real-time imaging and feature matching module:

[0058] Relying on GPU and parallel computing framework, a real-time imaging algorithm is applied to the preprocessed data to generate high-resolution SAR images of the ocean area. The calculation formula of the real-time imaging algorithm is: , clearly showing the locations of ships, islands, and coastlines, and synchronously calling the knowledge graph feature library. The ship's comprehensive features are compared with the templates in the knowledge graph (such as the feature templates of "cargo ship" and "fishing boat") through the multimodal matching algorithm. The calculation formula of the multimodal matching algorithm is: , output target candidate categories (such as "cargo ship", "tanker", "fishing boat").

[0059] Association reasoning and result optimization module:

[0060] Relying on GPU computing power, we integrate the relationship rules in the knowledge graph (such as "tankers usually sail between ports and oil fields") and process the candidate categories using the knowledge graph probabilistic reasoning model. The calculation formula of the knowledge graph probabilistic reasoning model is: For example, if a target is initially matched as an "oil tanker" and the knowledge graph shows that its route is close to ports and oil fields, the inference model will increase the confidence of the "oil tanker" category and ultimately output an accurate recognition result containing the ship category (such as "oil tanker"), location (such as the coordinates of a certain sea area), and confidence. High-confidence results (such as a "cargo ship" recognition result with a confidence level ≥ 90%) will be fed back to the knowledge graph for updating the feature library and relationship rules.

[0061] In summary, in marine ship monitoring, the system starts with data acquisition, obtains reliable data through polarization mode switching and verification mechanism, preprocessing and feature extraction highlight ship characteristics, high-speed data interaction ensures efficient data transmission, knowledge graph provides a basis for identification, real-time imaging and feature matching preliminarily determine candidate targets, associative reasoning optimizes results and improves knowledge graph, and each module works together, relying on relevant algorithms and functions to achieve real-time and accurate identification of marine ships.

[0062] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A high-computing SAR real-time imaging and target recognition system based on FPGA+GPU architecture, characterized by: The system includes: SAR echo data acquisition module: collects multi-band SAR echo data through the 16-bit AD / DA daughter card and FMC interface, and transmits it to the pre-processing module via the PCIe bus; Data preprocessing and feature extraction module: Based on FPGA, it performs noise reduction, filtering, and range compression on the raw data, simultaneously extracts texture, contour, and scattering features, and generates fused features through a multi-dimensional feature fusion algorithm with dynamic weight control; High-speed data interaction module: uses GTH high-speed interface and ETH Ethernet interface to realize bidirectional transmission of feature data, control instructions and status information between FPGA and GPU; Knowledge graph storage and update module: This module stores the knowledge graph constructed by SAR target feature attributes, category labels, and inter-target relationship data. It supports the dynamic addition of new features and relationship rules based on recognition results with a confidence level of ≥90%, and regularly clears feature data that has been stored for longer than a preset time limit. Real-time imaging and feature matching module: Relying on GPU and parallel computing framework to generate high-resolution SAR images, it calls the knowledge graph feature library and outputs target candidate categories through multimodal matching algorithm; Association reasoning and result optimization module: Integrates knowledge graph relationship rules to perform probabilistic reasoning on candidate categories, outputs recognition results containing target category, location and confidence, and feeds back results with confidence ≥ 90% to the knowledge graph storage and update module.

2. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the SAR echo data acquisition module, the AD / DA subcard supports dynamic switching of polarization modes. Real-time adjustment is achieved through FPGA configuration registers to achieve adaptive acquisition of HH, VV, HV, and VH polarization modes, and the multi-band switching response time is ≤100μs. When collecting data, parity check and CRC check are performed synchronously, and a unique check code is generated in combination with the timestamp.

3. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the data preprocessing and feature extraction module, the features of texture, contour, and scattering are extracted and fused by a multi-dimensional feature fusion extraction algorithm. The calculation formula of the multi-dimensional feature fusion extraction algorithm is: ,in is the polarization characteristic operator, is the scattering characteristic operator, is the texture feature operator, is a dynamic weight, which is adaptively controlled by the scene complexity calculated in real time by FPGA, and , For the moment The multi-dimensional comprehensive features after extraction and fusion, Indicates time.

4. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the knowledge graph storage and update module, the specific steps of constructing a knowledge graph based on SAR target feature attributes, category labels, and inter-target relationship data are as follows: (1) Data normalization: receiving SAR target characteristics, categories, and relationship data and standardizing the format; (2) Entity point construction: Based on the category label, the target is abstracted into an entity node and associated with feature attributes; (3) Relationship Edges: Analyze the relationship between targets, build directed edges between entity nodes, and mark the relationship type; (4) Knowledge fusion: merge duplicate entities and resolve relationship conflicts; (5) Rule deduction: Inferring implicit relationships based on existing knowledge and extracting high-frequency pattern generation rules; (6) Storage index: Store entities, relationships, and rules in a graph database format, establish classification indexes, and precipitate metadata.

5. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the knowledge graph storage and update module, features and relationships are dynamically supplemented. When the recognition result confidence level is ≥90%, the feature supplement process is triggered: the multi-dimensional feature vector of the newly recognized target is extracted. , Represent the characteristic components of different dimensions of the target, and calculate The similarity with each template in the existing feature library is as follows: , represent and The quantitative results of the similarity calculated between is the similarity calculation function, It represents the newly collected and newly input feature data used to participate in similarity calculation. It is the pre-stored feature data used to compare with the new feature. , then Add it to the library as a new feature template; when the recognition result reveals a new relationship, trigger the relationship supplement process: calculate the new relationship The frequency of occurrence of is: ,when hour, For this relationship The number of occurrences, Target The total number of related relationships will be Added to the knowledge base as a new relationship rule, Is a new relationship, Rel is a target object The related relationship function, is the target object in the SAR image.

6. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 5 is characterized in that: In the knowledge graph storage and update module, the confidence calculation formula is: ,in is the target The recognition confidence of is the posterior probability of knowledge graph association reasoning, is the similarity score between the target feature and the knowledge graph template, with a value range of , is the scene consistency score, is the dynamic weight coefficient, and , determined by the complexity of the current scene.

7. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the real-time imaging and feature matching module, a real-time imaging algorithm is used to generate a high-resolution SAR image. The calculation formula of the real-time imaging algorithm is: ,in yes High-resolution SAR images generated at all times, is a random sampling matrix, yes The square of the norm, It is the echo data after FPGA preprocessing. is the total variation regularization term, is the regularization coefficient, Represents the SAR image to be solved.

8. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the real-time imaging and feature matching module, a multimodal matching algorithm is used to compare the underlying features with the template and output the target candidate category. The calculation formula of the multimodal matching algorithm is: , is the matching score between the target feature and the template, is the underlying feature vector to be matched, is the target template feature in the knowledge graph, is a multimodal feature matching operator, is the association weight between the feature and the knowledge graph, represents the knowledge graph, is the distance between the feature and the template, It is a distance-based attenuation factor calculated for all target templates in the knowledge graph. , filter the top scorers The category of the item is used as a candidate category .

9. The high-computing-power SAR real-time imaging and target recognition system based on FPGA+GPU architecture according to claim 1 is characterized in that: In the association reasoning and result optimization module, the knowledge graph probability reasoning model is constructed and the calculation formula is: ,in: is the optimal target category for final reasoning, is a candidate category, It is the initial matching probability based on the underlying features, It is a set of relationship rules related to the target in the knowledge graph. is the rule weight, which is dynamically adjusted by the rule confidence and scene complexity. is the target and rules degree of correlation.

Images